【机器学习】线性回归算法应用

Ex2 实现房价预测的线性回归算法

Ex3 通过设置不同较大、合适和较小的学习率来对比不同的实验结果。

实验一代码：
 

# -*- coding: utf-8 -*-
import numpy as np
from math import e

# 迭代函数. 收敛性由定理2保证.
def Phi(x):
    return (6-x**3-e**x/2.0)/5.0

def main():
        x = np.array([1]) #x0
        x = np.append( x, Phi(x[0]) ) #x1
        print(f"x[0] = {x[0]:.8f}")        
        print(f"x[1] = {x[1]:.8f}, 误差为{np.abs(x[1]-x[0]):.8f}")        
        k=1 
        while( np.abs(x[k]-x[k-1]) > 1e-4):        
            x = np.append( x, Phi(x[k]) ) #x_k+1    
            print(f"x[{k+1:d}] = {x[k+1]:.8f}, 误差为{np.abs(x[k+1]-x[k]):.8f}")
            k+=1;         
            print(f"根为x[{k:d}] = {x[k]:.8f}")
            
if __name__ == "__main__":
    main()

实验二代码：

import pandas as pd
import matplotlib.pyplot as plt
raw_data = pd.read_csv(r'D:\研一\机器学习\kc_house_data.csv')
raw_data                   #查看原数据  
raw_data.info()              #查看数据特征类型
raw_data.duplicated().sum()   #检查是否有重复特征数据

X = raw_data.drop(['id', 'date', 'price'], axis=1)
y = raw_data['price']
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=1026)

from sklearn.preprocessing import StandardScaler
sc = StandardScaler()
sc.fit(X_train)
X_train= sc.transform(X_train)
X_test = sc.transform(X_test)

import numpy as np
# 计算损失函数值
def J(X_b,y,theta):
    try:
        return np.sum((X_b.dot(theta) - y) ** 2) / 2 * len(y)
    except:
        return float('inf')
    
# 计算梯度（针对一组theta参数的偏导数）
def dJ(X_b,y,theta):
    return X_b.T.dot(X_b.dot(theta) - y) / len(y)

# 以批量梯度下降法为例
def BGD(X_b,y,initial_theta,eta=0.01,iters=1e4,epsilon=1e-4):
    theta = initial_theta   # 将参数的初始值赋给theta变量
    curr_iter = 1
    while curr_iter < iters:  # 判断迭代次数
        gradient = dJ(X_b,y,theta)  # 计算当前theta对应的梯度
        last_theta = theta  # 先保存theta的旧值
        theta = theta - eta * gradient  # 更新theta
        cost_value = J(X_b,y,theta)  # 计算更新后的theta对应的损失函数值
        last_cost_value = J(X_b,y,last_theta)  # 计算更新前的theta对应的损失函数值
        # 如果两次损失函数的值相差的非常小，则认为损失函数已经最小了
        if abs(cost_value - last_cost_value) < epsilon:
            break
        curr_iter += 1  # 每次迭代完毕，迭代次数加1
        return theta,cost_value  # 返回最佳的theta和对应的损失函数值
# 拼接训练样本的X_b
X_b = np.hstack([np.ones((len(X_train),1)),X_train])

# 初始化theta
initial_theta = np.random.random(size=(X_b.shape[1]))
theta,cost_value = BGD(X_b,y_train,initial_theta,iters=1e5)

# 拼接测试样本的X_b
X_b_test = np.hstack([np.ones((len(X_test),1)),X_test])

# 在测试集上预测
y_predict = X_b_test.dot(theta)
from sklearn.metrics import mean_squared_error
r2 = 1 - mean_squared_error(y_test,y_predict) / np.var(y_test)
print("在测试集上的R^2为 ",r2)

from sklearn.linear_model import LinearRegression
linreg=LinearRegression()
linreg.fit(X_train,y_train)

from sklearn import metrics
y_train_pred = linreg.predict(X_train)
y_test_pred = linreg.predict(X_test)
train_err = metrics.mean_squared_error(y_train, y_train_pred) 
test_err = metrics.mean_squared_error(y_test, y_test_pred) 
print( 'Train_RMSE分别是: {:.2f}'.format(train_err) )
print( 'Test_RMSE分别是: {:.2f}'.format(test_err) )
predict_score = linreg.score(X_test,y_test)
train_score = linreg.score(X_train,y_train)
print('训练集上的正确率为:{:.2f} '.format(train_score) )
print('测试集上的正确率为:{:.2f} '.format(predict_score) )

# 输出模型测试集上的效果
%matplotlib inline
import matplotlib.pyplot as plt
fig = plt.figure(figsize=(8,5))
plt.plot(y_test, label='real')
plt.plot(y_test_pred, label='predict')
plt.legend()
plt.show()
new_X = np.array([[3,1,1110,5550,1,1,0,3,8,1110,0,1888,0,98178,47,-122,1555,8888]])
print("自己希望购买的房屋数据：",new_X)
new_X = sc.transform(new_X)
print("预测希望购买的价格为：{}".format(linreg.predict(new_X))) 

from sklearn.model_selection import GridSearchCV
model_select=LinearRegression()
parameters={'fit_intercept':['True','False'],'copy_X':['True','False'],'n_jobs':[1,2,3]}
model_gs=GridSearchCV(estimator=model_select,param_grid=parameters,verbose=3)
model_gs.fit(X_train,y_train)
print("最优参数：",model_gs.best_estimator_.get_params())
print("相关系数矩阵为:\n")
print(theta)

实验3代码：

X=raw_data.drop(['id','date','price','sqft_lot15','sqft_living15','long','lat','zipcode','yr_built','sqft_basement'],axis=1)
y=raw_data['price']
X_train,X_test,y_train,y_test=train_test_split(X,y,test_size=0.3,random_state=1026)
sc.fit(X_train)
X_train=sc.transform(X_train)
X_test=sc.transform(X_test)
# 拼接训练样本的X_b
X_b = np.hstack([np.ones((len(X_train),1)),X_train])
# 初始化theta
initial_theta = np.random.random(size=(X_b.shape[1]))
# 学习率为0.1
theta,cost_value = BGD(X_b,y_train,initial_theta,eta=0.1,iters=1e5)

# 拼接测试样本的X_b
X_b_test = np.hstack([np.ones((len(X_test),1)),X_test])
# 在测试集上预测
y_predict = X_b_test.dot(theta)
from sklearn.metrics import mean_squared_error
 
r2 = 1 - mean_squared_error(y_test,y_predict) / np.var(y_test)
print("在测试集上的R^2为：",r2)

# 拼接训练样本的X_b
X_b = np.hstack([np.ones((len(X_train),1)),X_train])
# 初始化theta
initial_theta = np.random.random(size=(X_b.shape[1]))
# 学习率为0.01
theta,cost_value = BGD(X_b,y_train,initial_theta,eta=0.01,iters=1e5)

# 拼接测试样本的X_b
X_b_test = np.hstack([np.ones((len(X_test),1)),X_test])
# 在测试集上预测
y_predict = X_b_test.dot(theta)
from sklearn.metrics import mean_squared_error
 
r2 = 1 - mean_squared_error(y_test,y_predict) / np.var(y_test)
print("在测试集上的R^2为：",r2)

总结：

  通过本次实验，我深入学习了如何用迭代法去求解方程式，并且实现了房价预测的线性回归方程，实践了样本数据分析与处理、划分训练集和验证集、对特征进行归一化、线性回归算法建模及分析和特征选择等过程。